Terrain Quantization

[bagnell]

Opening for early review.There will be a lot of reorganization and generalization for other geometry rendered RTC.

• Add quantized mesh skirts
• Fix compression with normals
• Add compression to meshes created from heightmaps
• Remove all of the uniforms. Only a matrix multiply is needed to undo the scale and bias.
• Update picking commands
• Update ray-terrain mesh intersection tests.
• Performance
• Tests
• Doc

[pjcozzi]

@mlimper this work is inspired by our conversation at SIGGRAPH about WEB3D_quantized_attributes. We are also going to implement that glTF extension.

[pjcozzi]

Should this be @private?

[pjcozzi]

Is this technically a breaking change since TerrainMesh is public?

[pjcozzi]

Looks good so far. Reasonably clean.

[bagnell]

Reasonably clean.

I moved all of the quantization and packing to TerrainEncoding I was hoping it could be more general, but its terrain specific only because of the height component.

[mlimper]

@mlimper this work is inspired by our conversation at SIGGRAPH about WEB3D_quantized_attributes. We are also going to implement that glTF extension.

Great, happy to hear that. Let me know if you discover things that could be improved in the extension.

[bagnell]

@pjcozzi This is ready for another review.

[bagnell]

For the same views as in the fog Sandcastle example:

memory	high alt	low alt
compressed CPU	224,368K	231,200K
uncompressed CPU	428,552K	426,096K
compressed GPU	216,600K	189,776K
uncompressed GPU	252,024K	192,020K

[pjcozzi]

Update CHANGES.md for 1.17 including the removed classes (that we'll deprecate in 1.16).

[pjcozzi]

For the same views as in the fog Sandcastle example:

"uncompressed" is before, and "compressed" is after, right?

The CPU results are, well, amazing.

Do you think the low altitude GPU results are suspect since there is such a small difference?

[pjcozzi]

Seems like we get a lot more cracking. Zoom into San Francisco. See the dotted black lines in the water?

[pjcozzi]

That is with terrain.

[pjcozzi]

Here is just off the coast in South Beach without terrain:

[pjcozzi]

If we can't track this down, worse case, I guess we can render the boundary (and/or perhaps skirts) with full precision. Or maybe we use double the required precision (if it turns out to still be smaller overall) so that adjacent tile boundaries at the same LOD are guaranteed to match up. Would need to think more about adjacent tiles at different LODs.

[bagnell]

The CPU results are, well, amazing.

They are a little better than I expected but in the same area since the number of bits per vertex was reduced by half.

Do you think the low altitude GPU results are suspect since there is such a small difference?

I think the GPU results are much closer because the textures are stored there and aren't kept on the CPU.

[pjcozzi]

I think the GPU results are much closer because the textures are stored there and aren't kept on the CPU.

Ah yes. You could also test with Natural Earth to see, but we would not publish those numbers since they won't be a typical case.

[pjcozzi]

The boundary artifact is the only issue. This is ready (post 1.16) otherwise.

[bagnell]

Here is just off the coast in South Beach without terrain

That is just the imagery. The same thing happens in master.

[pjcozzi]

Here is just off the coast in South Beach without terrain

That is just the imagery. The same thing happens in master.

Ah, OK. Can you reproduce the issue in SF?

[bagnell]

Fixes #3265.

[bagnell]

@pjcozzi The cracking issue with terrain was fixed. This is ready for another look.

[pjcozzi]

Please don't forget the blog post and final memory numbers.

[pjcozzi]

Looks good!